XXXXXXXXXXX was looking to improve the quality control of their production plant and both detect and grade their products into high grade, medium grade and unusable categories.
The plant was semi automated, with several steps including the quality control being handled manually.
The process was both slow, creating a bottleneck before the packaging process, and prone to human errors.
Batches of products being returned due to failure in meeting quality standards.
Batches where high grade products were sold in medium or low grade pricing due to incorrect classification.
Process more products per hour using a much smaller workforce
Produce more accurate results to reduce product return and incorrect pricing
We determined that the objectives could be achieved by performing the following operations:
An anomaly detection would be necessary to detect foreign objects or aspects of the production which needs to be completely avoided
An outlier detection would be necessary to help in grading, and detect a degree of presence or absence of features that are more, or less desired within the final product.
It was needed to point out the location of the defect for possible human verification and as a way to detect problems in accuracy with the detection mechanism to facilitate retraining and improvement.
A robotic loading/ offloading and conveyor belt mechanism connected with the detection system would speed up the process.
While in the implementation phase, we encountered some practical issues that required resolution.
Due to the factory setting, the image capturing mechanism is not ideal, and there is limited scope for improvement.
Sometimes defects or anomalies are camouflaged within the products themselves, as they are multi coloured.
Sometimes ‘fake’ defects or anomalies are recorded within the products due to reflections from surroundings and lighting conditions.
Overall the dataset was quite diverse, and there were relatively small margins between the grading of “great”,”usable” and “faulty”.
We introduced an enclosed ‘light box’ structure within the production pipeline where the images would be captured within an enclosed area whose ambient color can be controlled by using white-lined bulbs and software controlled RGB LED lights.
This allowed the system to protect the product from artificial artifacts or anomalies being registered due to lighting or reflections.
The system also allowed us to work with multiple images taken in different ambient colors, which solved the defect camouflage problem by giving our product outline detection, separation and defect outline detection algorithms multiple sets of data whose results could be cross referenced and used for better accuracy.
Preprocessing the images and removing the background using HSV color model and applying gaussian blur for removal of insignificant pixel level artifacts or noise which are not of significant size
f-AnoGAN with structural similarity index measure (SSIM) for anomaly localization
Speed of workflow was increased by 10x, from one product being inspected in 5 mins, to one product being accurately inspected in 30 seconds.
Product Rejection rate dropped from 25% to 10% and instances of improper grading was reduced to 15%.
Python
OpenCV
PyTorch
Scikit-Learn
TensorFlow
Keras
